Is anyone working on conversation compaction?

In our chat app "harness" we recursively generate summaries, L1 → L2 → L3. L1 summaries are more factual extraction than coherence, then get rolled into a more storytelling L2. Then we keep a tail of always ~10 raw messages with timestamps. We're not coding so this keeps context really clean like 5-

中文全文翻译正在补齐，可先查看下方英文原文。

In our chat app "harness" we recursively generate summaries, L1 → L2 → L3. L1 summaries are more factual extraction than coherence, then get rolled into a more storytelling L2. Then we keep a tail of always ~10 raw messages with timestamps. We're not coding so this keeps context really clean like 5-